Towards a Generalizable Machine-Learned Potential for Metal-Organic Frameworks

Machine-learned Potentials (MLPs) have transformed the field of molecular simulations by scaling `quantum-accurate' potentials to linear time complexity. Yet, while they provide a more accurate reproduction of structural properties as compared to empirical force fields, it is still computationally costly to generate their training dataset from \textit{ab initio} calculations. However, in the current literature, one MLP model is always specifically developed and employed for a specific system, and it is unexplored if one `general' MLP can be developed for a wide variety of structures. Herein, by leveraging upon data-efficient equivariant MLPs, we demonstrate the possibility of developing a general MLP for nearly 3,000 Zn-based Metal-Organic Frameworks (MOFs). After generating a training dataset comprising augmented structures generated from DFT-optimized ones, we validated the trained MLP's reliability in predicting accurate forces and energies when evaluated on the test set that comprises chemically distinct MOF structures. Despite incurring slightly higher errors on structures containing rare chemical moieties, the general MLP can still reliably reproduce thermal and bulk properties for a large sample of Zn-based MOFs. Crucially, by developing one MLP for many structures, the computational cost of developing MLP(s) for high-throughput screening is potentially reduced by a few orders of magnitude. This enables us to predict `quantum-accurate' properties for notable Zn-MOFs that were previously un-investigated via expensive theoretical calculations. More broadly, our contribution facilitates the development of general MLPs to accelerate chemical discoveries among other systems of interest at a fraction of the computational cost.